- All puzzles except 4 have lines
From a quick glance, text in these puzzles seems to mostly be ontop of lines on the pages (not free floating?) - We can then tell (most of the time) if the text on a piece is right side up or upside down based on the text’s location relative to the nearest line
- We can align all the pieces by rotating them until the lines become horizontal 3d Edge cuts
- The edge cuts are 3-dimensional in nature (not straight vertical slices). How do we properly account when trying to match up pieces based on edges? (since we don’t have a 3d model of each piece)
- We can see the edges blend out into the background gradually, so maybe we can make a guess as to how thick the edge is by looking at the ratio of the shard’s color to the background color at that point??
- Text based clustering
- For some of the images (1, 2, and 4), it seems that there are not that many text pieces compared to the overall number of pieces
- For these puzzles, maybe it makes sense to just filter out the text pieces and do manual assembly?
- The text in puzzle 3 seems to be much bigger than the shards, so any text-based clustering algorithm for this puzzle would probably be different than for the rest (assembly of each letter first before assembling words)